System and Method to Alter an Image

1. A computer-implemented method of altering an input image to provide an interpretability of a trained classifier, the method comprising the following steps: accessing the trained classifier and the input image, the trained classifier including multiple filters; applying the trained classifier to the input image to obtain a source classification; obtaining an average filter activation difference including a measure for a difference in filter activations of a first set of images having the source classification and a second set of images having a target classification; and determining an altered image for the input image, the determining of the altered image including optimizing at least: a classification score of the altered image to reduce a distance between a classification of the altered image and the target classification, and a similarity score to reduce a distance between the average filter activation difference and the measure for a difference in filter activations between the input image and the altered image.

2. The method of claim 1, further comprising the following step: displaying the altered image and/or a difference-image between the input image and the altered image.

3. The method of claim 1, further comprising the following step: retraining the trained classifier at least on the altered image.

4. The method of claim 1, wherein the trained classifier is trained through supervised learning on a training set of images and corresponding classifications, and wherein the first set of images and/or the second set of images are taken from the training set.

5. The method of claim 1, wherein the trained classifier includes at least one layer including multiple convolutional filters producing the filter activations.

6. The method of claim 1, wherein the accessing of the input image includes receiving the input image from an image sensor.

7. The method of claim 1, wherein the input image is obtained from an image sensor employed in a manufacturing line and the trained classifier is configured at least for ok/not-ok classification, and wherein further processing in the manufacturing line being dependent on the ok/not-ok classification.

8. The method of claim 1, further comprising the following steps: obtaining a first multiple of altered images from multiple input images and/or multiple target classifications; selecting from the first multiple of altered images a second multiple of altered images for which the trained classifier produces an incorrect classification and obtaining corresponding corrected classifications; and retraining the trained classifier at least on the second multiple of altered images and the corrected classifications.

9. The method of claim 1, wherein the obtaining of the average filter activation difference includes applying the measure to an averaged difference between filter activations in the trained classifier for the first set of images and filter activations in the trained classifier for the second set of images.

10. The method of claim 1, further comprising the following step: obtaining difference vectors indicating a difference in averaged difference between filter activations in the trained classifier for the first set of images and filter activations in the trained classifier for the second set of images, the measure including applying a kernel function to each combination of difference vectors.

11. The method of claim 1, wherein the measure includes computing a Gram matrix.

12. The method of claim 1, wherein the optimizing includes minimizing a loss term, the loss term including LG(x′)=Σl∥Gl(Dl(Cx,Cy))−Gl(Dl(x,x′))∥, wherein x′ denotes the altered image, Gl denotes the Gram operator to compute a Gram matrix from a set of vectors, Dl(Cx,Cy)=FCxl−FCyl, wherein Cx is the first set of images, Cy is the second set of images, FCxl and FCyl are averaged filter activations at layer l for the first set Cx and second set Cy respectively, Dl(x, x′) denotes the difference in filter activations at layer l for the input image x and the altered image x′.

13. The method of claim 1, wherein the optimizing includes minimizing an uncertainty estimation of the altered image.

14. The method of claim 1, wherein the trained classifier is configured to selectively export filter activations when applied to the input image.

15. An image processing system for altering an input image to provide an interpretability of a trained classifier, the image processing system comprising: an interface configured for accessing the trained classifier and the input image, the trained classifier including multiple filters; and a processor subsystem configured to: apply the trained classifier to the input image to obtain a source classification, obtain an average filter activation difference including a measure for a difference in filter activations of a first set of images having the source classification and a second set of images having a target classification, determine an altered image for the input image, the determination of the altered image including optimizing at least: a classification score of the altered image to reduce a distance between a classification of the altered image and the target classification, and a similarity score to reduce a distance between the average filter activation difference and the measure for a difference in filter activations between the input image and the altered image.

16. A non-transitory computer readable medium on which is stored data representing instructions for altering an input image to provide an interpretability of a trained classifier, the instructions, when executed by a processor system, causing the processor system to perform the following steps: accessing the trained classifier and the input image, the trained classifier including multiple filters; applying the trained classifier to the input image to obtain a source classification; obtaining an average filter activation difference including a measure for a difference in filter activations of a first set of images having the source classification and a second set of images having a target classification; and determining an altered image for the input image, the determining of the altered image including optimizing at least: a classification score of the altered image to reduce a distance between a classification of the altered image and the target classification, and, a similarity score to reduce a distance between the average filter activation difference and the measure for a difference in filter activations between the input image and the altered image.